[m-dev.] tail call to loop optimisation for low-level grades

[Peter Wang]

On 2008-07-01, Zoltan Somogyi <zs at csse.unimelb.edu.au> wrote:
> On 27-Jun-2008, Peter Wang <novalazy at gmail.com> wrote:
> > I was looking at why the asm_fast.gc implementation of string.hash is
> > about 85% times slower than the C version.  The loop looks like this,
> > after some cleanup:
> > 
> >     MR_def_static(shash__unchecked_hash_2_5_0)
> > 	    if (MR_r2 >= MR_r3) {
> > 		MR_GOTO_LAB(shash__unchecked_hash_2_5_0_i2);
> > 	    }
> > 	    {
> > 		MR_String Str = (MR_String) MR_r1;
> > 		MR_Word MR_tempr1 = Str[MR_r2];
> > 
> > 		MR_r2 = MR_r2 + 1;
> > 		MR_r4 = (MR_r4 ^ (MR_r4 << 5)) ^ MR_tempr1;
> > 		MR_np_localtailcall(shash__unchecked_hash_2_5_0);
> > 	    }
> 
> Your cleanup shows the C << operator, whereas the Mercury source in string.m
> uses the Mercury << operator, which does bounds checks. Did you replace
> << with unchecked_left_shift in your copy of string.m?

Yes.  That's what it should be if string.hash is to match the
MR_hash_string() macro.

BTW, for this example, it turns out that the slowdown is because r4 is a
fake register.  If you just exchange r1 and r4 before entering the loop,
you get the same (or similar, I forget) speedup.  So maybe that is a
better way to go?

Peter

--------------------------------------------------------------------------
mercury-developers mailing list
Post messages to:       mercury-developers at csse.unimelb.edu.au
Administrative Queries: owner-mercury-developers at csse.unimelb.edu.au
Subscriptions:          mercury-developers-request at csse.unimelb.edu.au
--------------------------------------------------------------------------